Thesys Open-Weights OUI-1 Hits 71.7% on Generative UI Benchmark

Thesys’s Engineering Team published OUI-1 on Tuesday, September 8, 2026: an open-weight DiffusionGemma finetune that writes user interfaces in openui-lang and scores 71.7% on the Generative UI Benchmark—5.5× the untuned DiffusionGemma base at 13.0%. The model is a 26B-A4B design (26 billion parameters, 4 billion active), runnable on a consumer GPU such as an RTX 5090 at FP8, with weights on Hugging Face at thesysdev/OUI-1 under the Gemma Terms of Use.

OUI-1 targets agent-driven interfaces that must generate quickly, parse cleanly, and fit on-device hardware. OpenUI Lang is positioned as a compact declarative format (the launch post cites up to 67% fewer tokens than JSON) that can stream as the model generates. Thesys chose DiffusionGemma because it writes 256-token blocks in parallel rather than one autoregressive token at a time; Google’s cited speeds for the base include over 1,000 tokens per second on an H100 and over 700 on an RTX 5090.

Training: SFT, then parser-rewarded self-distillation

Training ran in stages. Supervised fine-tuning on OpenUI Lang first raised the single-library benchmark score from 13.0% to 28.8%, but schema and wiring errors traded off and generation slowed. Self-distillation then used the OpenUI Lang parser as a verifiable reward: generate programs, keep or repair parser-accepted near-misses, and retrain. That loop cut both error types and recovered speed. Thesys repeated the SFT-plus-self-distillation recipe across 27 component libraries to generalize beyond the benchmark library.

Benchmark and AppLess results

On the Generative UI Benchmark (46 briefs, four generations each; HF card reports 132/184 for OUI-1 vs 24/184 for the base), OUI-1’s 71.7% beats every listed open model with up to 31B active parameters except Qwen3.8 27B at 78.8%. Thesys notes Qwen3.8 is dense 27B on every token versus OUI-1’s 4B active; nothing at or below 4B active scored higher on that table. Gemma 4 31B sits at 46.7% and Gemma 4 26B-A4B at 29.9%. On sixty unseen AppLess asks, OUI-1 produced 55/60 valid outputs versus DiffusionGemma’s 23/60.

The Hugging Face card lists base google/diffusiongemma-26B-A4B-it, LoRA merged into bf16 safetensors, 16,384-token context as served, and vLLM ≥0.24 FP8 serving guidance. Outputs are intended for openui-lang validation and library rendering—not general chat.

Primary sources for this brief are the OpenUI / Thesys launch post (https://www.openui.com/blog/oui-1) and the Hugging Face thesysdev/OUI-1 model card (https://huggingface.co/thesysdev/OUI-1).